Presentation by Dr. Cliff Click, Jr. Mention Java performance to a C hacker, or vice versa, and a flame war will surely ensue. The Web is full of broken benchmarks and crazy claims about Java and C performance. This session will aim to give a fair(er) comparison between the languages, striving to give a balanced view of each language's various strengths and weaknesses. It will also point out what's broken about many of the Java-versus-C Websites, so when you come across one, you can see the flaws and know that the Website isn't telling you what it (generally) claims to be telling you. (It's surely telling you "something," but almost just as surely is "not realistically" telling you why X is better than Y).
2. Java vs C/C++
●
"I declare a Flamewar!!!!"
●
Lots of noise & heat
●
Not many facts
●
Lots of obvious mistakes being made
●
Situation is more subtle than expected
●
This is my attempt to clarify the situation
3. C/C++ Beats Java
●
Very small footprint – under 300KB
●
●
Very deterministic or fast (re)boot times –
●
●
e.g. engine controllers, pacemakers
Very big problems: Fortran optimizations
●
●
e.g. Embedded controllers, cars, clocks
Array reshaping & tiling for cache
Value types - Complex, Point
●
e.g. Overhead costs of 1b objects
●
vs array-of-doubles
4. C/C++ Beats Java
●
Direct Machine Access
●
e.g. OS's (special ops, registers), device drivers
–
Hard to do in Java (i.e. JavaOS effort)
●
●
●
AAA Games / First Person Shooter Games
Maxine Java-in-Java might be a counter-example
Direct Code-Generation
●
gnu "asm"
●
Write bits to buffer & exec
–
●
'sort' inner loop key-compare
Interpreters
5. C++ Beats Java
●
Destructors vs finalizers
●
Destructors are reliable out-of-language cleanup
●
Finalizers will "eventually" run
–
–
●
But maybe after running out of e.g. file handles
So weird force-GC-cycle hooks to force cleanup
Destructors vs & try/finally
●
Destructors are reliable exit-scope action
●
try/finally requires adding explicit exit-scope-action
–
–
For each new enter-scope-action
Maintenance mess
6. Java Beats C/C++
●
Most Programs - profiling pays off
●
●
All JIT systems profile at least some
●
●
But nobody bothers for C/C++, too hard
More profiling added as systems mature
Very Large Programs >1MLOC
●
Large program tool chain is better
●
A lot more 1MLOC Java apps than C
7. Java Beats C/C++
●
GC is easier to get right than malloc/free
●
●
●
Faster time-to-market
Why so many variations on Regions, Arenas,
Resource Areas? Basically hand-rolled GC...
GC is efficient
●
●
●
Parallel, concurrent
Good locality, fragmentation
GC allows concurrent algorithms
●
Trivially track shared memory lifetimes
●
Fundamental change, can't "fake it"
8. Java Beats C/C++
●
Single CPU speed stalled
●
●
Bigger problem => parallel solution
Better multi-threading support
●
Real Memory Model - synchronized, volatile
●
Threads are built-in
●
Large multi-threaded library base
–
●
●
JDK Concurrent Collections
GC vs concurrent malloc/free
Tools for parallel coding, debugging
9. Libraries
●
Vast Java Library collection
●
●
Can COTS many many problems
Downside: too many 3rd party libraries
●
●
C Mentality: build before download
●
Too many layers of Java crap
●
●
Java Mentality: download from web, don't build
Nobody knows what's going on
Application plagued by failures
no one understands
10. Claims C-beats-Java
But I Dont Think So
●
Most modest sized programs
●
●
Fast enough is fast enough
16bit chars vs 8bit chars
●
●
●
Lots of noise here (and lots of optimizations)
Rarely makes a difference in practice
Raw small benchmark speed
●
Usually I don't care
–
●
"C gets more BogoMips so it's better!"
OR broken testing methodology
–
"C makes a better WebServer because printf is faster!"
11. Common Flaws
When Comparing
●
No Warmup
●
●
●
Only interesting for quick-reboot, e.g. Pacemakers
Most apps run for minutes to days
Basic timing errors
●
●
●
API reports in nanos
OS rounds to millis (or 10's of millis)
Caching Effects
●
CPU caches, OS-level, disk & network
●
DB cache, JIT/JVM level
vs
12. Common Flaws
When Comparing
●
Basic Broken Statistics
●
Run-once-and-report
●
No averages, std-devs
●
Throwing out "outliers"
●
Not accounting for "compile plan"
–
–
●
"Statistically rigorous Java performance evaluation"
"Producing wrong data without doing anything obviously
wrong!"
Flags, version-numbers, env-factors all matter
●
"java" not same as "java -client" or "java -server"
●
Some JDK versions have 30% faster XML parsing
13. Common Flaws
When Comparing
●
Varying Datasets or Constant-time workloads
●
●
Have seen cycles-per-work-unit vary by 10x
Claiming X but testing Y
●
●
SpecJBB: claims middleware test but is GC test
●
●
209_db: claims DB test but is shell-sort test
Lots more here
Not comparing same program
●
e.g. Debian language shootout
–
http://shootout.alioth.debian.org
14. Commonly Mentioned
Non-Issues
●
Stack Allocation "Does So" beat GC
●
●
Does Not. You got evidence?
I got evidence of non-issue...
Java has lots of casts
●
But they are basically free
–
●
Virtual & Interface calls are slow
●
●
load/compare/branch, roughly 1 clock
And basically never taken (inline-cache)
C# curious? I dunno, I don't track Microsoft
15. Java-vs-C Examples
●
Examples limited to what I can fit on slides
●
In-Real-Life never get apples-to-apples
●
Programs either very small
●
Or new re-implementation
●
●
Generally better written 2nd go-round
Or extremely bad (mis)use of language features
16. Example: String Hash
●
Java tied vs GCC -O2
int h=0;
for( int i=0; i<len; i++ )
h = 31*h+str[i];
return h;
Here I ran it on a new X86 for 100 million loops:
> a.out
100000000
100000000 hashes in 5.636 secs
> java str_hash 100000000
100000000 hashes in 5.745 secs
●
Key is loop unrolling
●
(i.e. JITs do all major compiler optimizations)
17. Sieve of Erathosthenes
●
Again tied
bool *sieve = new bool[max];
for (int i=0; i<max; i++) sieve[i] = true;
sieve[0] = sieve[1] = false;
int lim = (int)sqrt(max);
for (int n=2; n<lim; n++) {
if (sieve[n]) {
for (int j=2*n; j<max; j+=n)
sieve[j] = false;
}
}
I computed the primes up to 100million:
> a.out
100000000
100000000 primes in 1.568 secs
> java sieve 100000000
100000000 primes in 1.548 secs
18. Silly Example
●
Silly Example to Make a Point
int sum=0;
for (int i = 0; i < max; i++)
sum += x.val(); // virtual call
return sum;
Here I run it on the same X86:
> a.out
1000000000 0
1000000000 adds in 2.657 secs
> java vcall 1000000000 0
1000000000 adds in 0.0 secs
●
Zounds! Java is "infinitely" faster than C
??? what happened here ???
19. Silly Example Explained
●
Command-line flag picks 1 of 2 classes for 'x'
●
Type profiling at Runtime
●
Only 1 type loaded for 'x.val()' call
–
●
JIT makes the virtual call static, then inlines
–
●
"for( int i=0; i<max; i++ ) { sum += 7/*x.val*/; }"
Once inlined, JIT optimizes loop away
–
●
"int val() { return 7; }"
"sum += max*7;"
True virtual call at static compile-time
●
No chance for a static compiler to optimize
20. Why Silly Example Matters
●
Only 1 implementing class for interface
●
Common case for large Java programs
●
Single-implementor interfaces abound
●
Library calls with a zillion options
–
●
But only a single option choosen, etc
Can see 100+ classes collapsed this way
–
10K call-sites optimized, 1M calls/sec optimized
●
Major Optimization not possible without JIT'ing
●
Lots more cool JIT tricks to come...
21. Other Stuff That Matters
●
Other Things Also Matter
●
Existing infrastructure, libraries, time-to-market
●
Programmer training, mind set
–
Lots of Java programmers Out There
●
●
●
Legal issues – open source or man-rating
Reliability, stability, scalability
JVMs enabling new languages
●
Clojure, Scala, JRuby, Jython, many more
●
Much faster time-to-market
22. Summary
●
My Language is Faster!!!
●
Except when it's not
●
Ummm.... "fast" is not well-defined...
–
●
Other-things-matter more in many domains
●
●
MOOPS/sec? Faster than thy competitor?
Faster-to-market? Fits in my wrist watch?
If you got 500 X programmers, maybe should use X?
Each language is a clear winner
in some domains, neither going away soon
●
e.g. still room for trains in our auto-dominated world
23. Summary
●
Internet is a Great Amplifier
●
●
Of both the Good, the Bad AND the Ugly
Real issue: Need Sane Discourse
●
Lots of Screaming & Flames
–
–
●
People with strong opinions, different vested interests,
different experiences & goals
e.g. Do we even agree on what "faster" means?
Lots of Bad Science
–
–
Broken & missing statistical evidence
Misapplied testing, testing unrelated stuff
24. Summary
●
When the noise exceeds communication levels...
●
●
●
Back up, clarify, acknowlege each side has strengths
Chill out, think it through
Recognize a lack-of-evidence for what it is
●
●
●
yelling louder about what you do know doesn't help
Good testing helps (and bad testing hurts)
Realize "faster" has different meanings
–
–
–
Junior Engineer thinks "faster than the competition"
Manager thinks "faster to market"
Senior Engineer thinks "that brick wall is approaching fast!"